Talk at Theoretical Foundations of ML conference

Mateusz Półtorak
dodane przez: Mateusz Półtorak | Maj 24, 2019

On 11–15 February 2019, a part of our team participated in the Theoretical Foundations of Machine Learning conference organized by the Jagiellonian University in Kraków. These five days were filled with plenary lectures by speakers from such groups as DeepMind and Google AI, as well as shorter talks and posters presented by conference participants.

The most comprehensive lectures were dedicated to topics such as:

  • reinforcement learning methods used in DeepMind,
  • various aspects of training variational auto-encoders and GANs,
  • graph-based neural networks and variational inference in deep learning.
  • Interestingly, a lot of lectures described the usage of machine learning in medicine (for example drug discovery or breast cancer diagnosis).

I’ve really enjoyed Wojciech Czarnecki’s talk about reinforcement learning population-based methods. The presentation was well prepared and contained tons of useful knowledge for people who are not reinforcement learning experts. In another talk, Ilya Tolstikhin gave the audience a few good examples of problems when VAEs are better choice than GANs and vice versa. Another interesting lecture was given by Mohammad Emtiyaz Khan from RIKEN. I saw how passionate he is about machine learning, so his talk about neural-gradient descent for variational inference, which could be used to estimate the confidence or uncertainty in predictions, was truly enjoyable.

I too had the pleasure of giving a talk. I spoke about emotion recognition in speech recordings using convolutional neural networks and spectrograms. The key idea of my presentation was to promote machine learning systems that are able to analyse complex data without additional preprocessing done by humans. I illustrated the idea with the results of my own experiment, in which deep convolutional neural network VGG-19 was trained on spectrograms. The network achieved close to the state-of-the-art performance in classifying emotions in speech recordings.

I highly recommend attending the next edition of the TFML conference. It’s a great conference and the beautiful city of Kraków is worth visiting.

Mateusz Półtorak

Mateusz Półtorak

Mateusz Półtorak is an Associate Data Scientist in Pearson's AI Products & Solutions unit. He is interested in using neural networks for computer vision (e.g. sign language recognition) and natural language processing (in both speech and writing).
Informacja dotycząca plików cookies

Informujemy, iż w celu optymalizacji treści dostępnych w naszym serwisie, dostosowania ich do Państwa indywidualnych potrzeb korzystamy z informacji zapisanych za pomocą plików cookies na urządzeniach końcowych użytkowników. Pliki cookies użytkownik może kontrolować za pomocą ustawień swojej przeglądarki internetowej. Dalsze korzystanie z naszego serwisu internetowego, bez zmiany ustawień przeglądarki internetowej oznacza, iż użytkownik akceptuje stosowanie plików cookies.

Akceptuję